I now implemented a prototype for JDK-6824466: [1]
The goal I set here was to remove the old AccessorGenerator.
The good news:
- It works - all tier1 tests pass, except one that requires the old
implementation.
- A lot of code has been removed. And even more could be removed.
The bad news:
- I hit a performance regression [2].
The JMH tests were written by Peter Levart [3].
As much as I don't like the old AccessorGenerator, I have to say that it
is very optimized.
If you have any ideas on how the performance could be improved, or how I
could find that out where the performance drops, let me know.
- Johannes
PS.: I played around with adding some @Stable and @ForceInline
annotations, and got a big performance improvement for the const cases
(~9 ns/op). But that came with other performance regressions for the
non-const cases.
[1]: https://github.com/openjdk/jdk/pull/2457
[2]:
https://gist.github.com/DasBrain/09b37510955b14f0ad82313df8d8de37#gistcomment-3623292
[3]:
http://cr.openjdk.java.net/~plevart/jdk-dev/6824466_MHReflectionAccessors/
